Four backends (vLLM FP16, vLLM INT4, SGLang, llama.cpp CPU) with adaptive routing, prefix caching, quantization, and speculative decoding.
Headline results: INT4 fast path 3.5× lower ITL (9.3 ms/tok), LMCache MP-mode cross-restart KV persistence, n-gram spec decode 98% acceptance on repetition-heavy tasks.